Research map / Image generation / Fast diffusion sampling
Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning
Rogério Guimarães, Pietro Perona
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PSP seed pruning instead of Best-of-N at matched compute N=4 | SD v1.5 · GenEval prompts | GenEval score: 0.542 -> 0.574 | Improved |
| use PSP instead of Best-of-N at matched compute | SDXL · GenEval prompts | GenEval score: 0.629 -> 0.645 | Improved |
| use PSP instead of Best-of-N at matched compute | SD 3.5 · GenEval prompts | GenEval score: 0.747 -> 0.747 | No clear effect |
| use PSP instead of Best-of-N, human evaluation | SD 3.5 · GenEval prompts human eval | human prompt-alignment rate: 0.831 -> 0.841 | Improved |
| use PSP instead of Best-of-N, human evaluation | SDXL · GenEval prompts human eval | human prompt-alignment rate: 0.682 -> 0.713 | Improved |
| use PSP instead of BFS (reward vs prompt alignment) | SDXL · GenEval prompts | ImageReward / GenEval: IR 1.247 -> 1.224; GenEval 0.636 -> 0.645 | Mixed |
| apply PSP on DPO-finetuned backbone instead of Best-of-N | SD v1.5 DPO · GenEval prompts | GenEval score: 0.562 -> 0.593 | Improved |
| apply PSP on DPO-finetuned backbone instead of Best-of-N | SDXL DPO · GenEval prompts | GenEval score: 0.657 -> 0.667 | Improved |
| use PSP over prompt rewrites instead of Best-of-N with fixed seed | SD v1.5 · GenEval prompts (prompt rewrites) | GenEval score: 0.512 -> 0.533 | Improved |
| use PSP over prompt rewrites instead of Best-of-N with fixed seed | SDXL · GenEval prompts (prompt rewrites) | GenEval score: 0.598 -> 0.604 | Improved |
| use HPS guidance with PSP instead of Best-of-N | SD v1.5 · GenEval prompts (HPS guidance) | GenEval score: 0.525 -> 0.541 | Improved |
| tune pruning schedule by grid search instead of default halving schedule (IR guidance) (ablation) | PSP SDXL · GenEval prompts | GenEval score: 0.645 -> 0.641 | No clear effect |
Closest papers by meaning
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.93)
- Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing (NeurIPS 2025 · similarity 0.92)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models (NeurIPS 2025 · similarity 0.90)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.90)
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
